How to use agent mode in ChatGPT now that it is ChatGPT Work

Where has ChatGPT agent mode gone?

Agent mode went into ChatGPT Work in August 2026. OpenAI's help page for ChatGPT agent, read on 27 September 2026, opens with "ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables," and OpenAI Support repeated that on 25 September 2026 to a Pro subscriber whose Tools menu had lost the option in July. There was no release note. The same page still lists the old caps under the notice, 40 messages a month on Plus and 400 on Pro, which is why guides ranking for this question in September 2026 still tell you to open Tools and pick Agent mode.

The sequence, from OpenAI's own release notes and launch posts: Operator was announced as a research preview on 23 January 2025 and shut on 31 August 2025 after agent mode absorbed it on 17 July 2025. Atlas, the browser that carried agent mode on Plus and Pro from 21 October 2025, had its retirement announced on 9 July 2026 and stopped working on 9 August 2026. ChatGPT Work was introduced on that same 9 July, and on 25 August 2026 its cloud browser learned to sign in to websites, on Plus and Pro. Agent mode vanished from the menus in between, and as of September 2026 the help page is the only notice.

What you lost and what you kept. Agent mode ran a virtual computer with a visual browser, a text browser, a terminal and connectors, paused at logins so you could take over, and counted tasks by the month. Work keeps the cloud browser, the sign-in pause, the confirmation prompts and scheduled tasks, adds finished files (documents, spreadsheets, slides, Sites) as the default output, and drops the monthly count for a 5-hour window shared with Codex. What it does not keep is the button: there is no /agent command, and the browser starts itself.

How do you start an agent task in ChatGPT now?

You start it by switching the chat to Work, describing the outcome, and letting ChatGPT decide when to open its browser. OpenAI's cloud browser page says it plainly: "You do not need to select the cloud browser separately. ChatGPT chooses when to use it." The steps below are from the ChatGPT Work and cloud browser help pages as of September 2026.

1. On chatgpt.com or in the iOS or Android app, open the dropdown at the top of the screen and pick Work instead of Chat. In the desktop app, choose ChatGPT from the top-left menu, then Work on the toggle at the top of the page.
2. Describe the result you want, naming the website, the details and any constraints. A prompt that works in this shape: "Go to the Ontario DMV site, find the earliest licence renewal appointment within 20 km of M5V 2T6 on a weekday after 3pm, and prepare the booking for my approval. Do not submit anything."
3. Approve website access when ChatGPT asks. Settings > Cloud browser holds the default (Always ask, Auto approve, or Always allow, which OpenAI itself marks "not recommended") and per-site allow and block lists.
4. If the site needs a login, ChatGPT pauses and shows a secure sign-in form. Your username and password go straight to the remote browser; OpenAI says they "are not visible to the model" and are not stored. Two-factor codes go in the same way.
5. Confirm anything consequential in the chat. The help page: ChatGPT "is designed to request confirmation before actions that could be hard to reverse or create a financial, legal, account, or other real-world commitment, such as confirming a booking or making a payment."
6. Close the laptop if you like. The task runs on OpenAI's computer and pauses only for input, sign-in or confirmation. If it gets stuck, it sends a link to take over the cloud browser from your phone or computer.

Two other surfaces exist. The desktop app has a built-in browser you and ChatGPT share on screen (Cmd+Shift+B, or Ctrl+Shift+B on Windows, or address it as @Browser in a Work or Codex task); it asks before it uses a new site, confirms "sensitive actions such as submitting information, making a purchase, changing permissions, or deleting data", and "can't automate file uploads". Atlas, the third surface, stopped working on 9 August 2026 and does not open any more. On mobile the cloud browser is the whole story: Codex is not selectable on iOS or Android, and Work is.

Is agent mode available on the Free plan, and what does each plan get?

No. Agent mode was never on Free, and the browser that replaced it is not either: OpenAI's pricing page lists Work as included on every plan down to Free, but the cloud browser that does the agent's job "is available in ChatGPT Work on paid ChatGPT plans in supported regions, excluding Free and Go." What changed for paying plans is the unit. Agent mode counted tasks per month. Work counts tokens per 5-hour window, because "ChatGPT Work usage inside ChatGPT uses the same pricing, credits, and usage limits as Codex."

The old column is from the ChatGPT agent help page as of September 2026, the new one from OpenAI's Codex pricing page, which states "the estimates below show local messages per five-hour period" and adds a weekly cap on top. A short lookup on GPT-6 Luna is a different budget entirely: 350 to 3,000 messages per window on Plus. Astra, the model OpenAI says has the better visual judgment for checking a page, "can use your allowance faster than GPT-5.6 Sol." One timing wrinkle: the 5-hour limit was lifted from Plus for several weeks in summer 2026, then put back on 25 August 2026 while the two Pro tiers stayed without it for the following months, so a Pro subscriber in September 2026 sees only the weekly cap.

The practical arithmetic. A Plus subscriber who used all 40 agent tasks a month before now gets up to 45 Astra-level tasks per 5 hours, which is the same month of work in an afternoon if the tasks are short, and less than a month if each task runs 20 minutes of browsing on a large page, because tokens, not tasks, are what drain the window. ChatGPT Plus vs Pro has the rest of the two plans' limits side by side.

What does the agent do well, and where does it fail?

It does research and logged-in lookups well and fails at layout, CAPTCHAs and any site that blocks automated browsers. We did not run agent mode for this page: it was gone before drafting began, and we did not drive a paid seat through the cloud browser, so every outcome below is someone else's dated run or OpenAI's own number.

DataCamp ran five tasks on launch week, published 24 July 2025. It built a spreadsheet of UNESCO education data for 222 countries in 14 minutes, missing only a requested summary tab. It produced a 10-slide pitch deck in 13 minutes that the reviewer called professional and well structured but not polished enough for a client. It made a usable branded cover image. It checked 3,000 Lego parts on BrickLink and corrected three part names in about five minutes. Then the failures: an image collage took more than 75 minutes and came out wrong, the BrickLink wishlist never got created, and the reviewer's summary of the mouse work was "like watching a child use a computer." MacStories tried the new cloud browser on 26 August 2026: on an iPhone it signed in to a password site and an affiliate dashboard and read the numbers back, with the author solving a CAPTCHA by hand; on a Mac the same flow froze and never completed.

OpenAI's own numbers from the 17 July 2025 launch point the same way. BrowseComp, its hard-to-find-information benchmark: 68.9%, 17.4 points above deep research. SpreadsheetBench with direct file editing: 45.5% against 20.0% for Copilot in Excel. Slideshows, in OpenAI's words, "can sometimes feel rudimentary in its formatting and polish."

What it will not do is written down. The launch post: "ChatGPT is trained to actively refuse high-risk tasks such as bank transfers," sending email needs "Watch Mode" with you present, and purchases need explicit confirmation. The cloud browser page adds the wall you cannot argue with: "A website may block ChatGPT from accessing the site or completing your task, even if the same website works normally in your own browser," and "These restrictions are set by the website, not by ChatGPT." CAPTCHAs are the usual form of that wall, and OpenAI's advice is to take over or open the page yourself. Never paste a password, code or card number into the chat; the secure form exists so the model never sees them.

Connectors are the other half of the toolbox, and the other half of the risk. Work can read Gmail, GitHub, Slack and the rest through connected apps, and OpenAI's advice is to "prefer a connected app or plugin when it supports the task directly" because an API call is cheaper and safer than a browser. The risk is prompt injection: the help page's own example is a malicious comment on a page that tells the agent to fetch a password reset code from Gmail and send it away. Enable only the apps a task needs, and read what is MCP for how those app connections are wired underneath.

Which AI assistants have an agent mode in September 2026?

Four consumer assistants run a browser for you as of 27 September 2026: ChatGPT through Work, Claude through its Chrome extension and the Cowork built-in browser, Gemini through Spark, and Perplexity through the Comet Assistant. Every one of them is paid-only, and only Perplexity still prints a task count. The table is compiled from each vendor's help pages on that date, for individual plans.

Claude's safety page carries the one number in that column: Anthropic says its current configuration "reduces attack success rates to less than 0.08% against our internal testing" for prompt injection, and in the next breath that "the chances of an attack are still non-zero." Google calls Spark "an experimental feature in early development" that "can make mistakes and do unexpected things." OpenAI runs Work on GPT-6 Sol and GPT-6 Luna, two models that do not appear in a regular chat at all, and says GPT-6 Astra is the one to pick for visual judgment when it is offered.

Whizi, which publishes this page, has an Agents section in its model picker, and it is honest to say what those are: four presets on the Pro plan (Canadian CPA, Study Mentor, Writing Copilot, Math Tutor), each a fixed system prompt with a bigger token budget on the same base model, and the CPA one forces a web search on every turn, capped at seven searches. None of them opens a browser, clicks a button or keeps running after you close the tab, so for a DMV appointment or a logged-in utility account Whizi loses to every row in the table. What it offers instead is web search on by default and GPT, Claude and Gemini on one login (280+ models on Powerhouse, trial week from $0.99 on Starter), which covers the reading half of what people used agent mode for and none of the clicking half. What is an AI agent draws that line in detail.

The decision rule, then. Use Work's cloud browser when the task has a login, a form or a booking at the end and you can sit through a confirmation prompt; use a chat with search for anything you only need to read, because it costs a fraction of the window and cannot buy anything by mistake.